Scheduling Techniques for Reducing Processor Energy Use in MacOS Jacob R. Lorch and Alan Jay Smith Computer Science Division, EECS Department, UC Berkeley Berkeley, CA 94720-1776 Abstract a number of engineering users. We found that, depending on the machine and user, up to 18±34% of total power was at- The CPU is one of the major power consumers in a tributable to components whose power consumption could be portable computer, and considerable power can be saved by reduced by power management of the processor, i.e. the CPU turning off the CPU when it is not doing useful work. In Ap- and logic that could be turned off when the CPU was inactive. ple's MacOS, however, idle time is often converted to busy This high percentage, combined with our intuition that soft- waiting, and generally it is very hard to tell when no use- ware power management could be signi®cantly improved for ful computation is occurring. In this paper, we suggest sev- the processor, led us to conclude that the most important tar- eral heuristic techniques for identifying this condition, and get for further research in software power management was for temporarily putting the CPU in a low-power state. These the processor. techniques include turning off the processor when all pro- Many modern microprocessors have low-power states, in cesses are blocked, turning off the processor when processes which they consume little or no power. To take advantage of appear to be busy waiting, and extending real time process such low-power states, the operating system needs to direct sleep periods. We use trace-driven simulation, using proces- the processor to turn off (or down) when it is predicted that sor run interval traces, to evaluate the potential energy sav- the consequent savings in power will be worth the time and ings and performance impact. We ®nd that these techniques energy overhead of turning off and restarting. In this way, save considerable amounts of processor energy (as much as the goal of processor power management strategies is simi- 66%), while having very little performance impact (less than lar to that of hard disks [3, 7]. Some strategies for making 2% increase in run time). Implementing the proposed strate- these predictions are described by Srivastava et al. [15] Un- gies should increase battery lifetime by approximately 20% like disks, however, the delay and energy cost for a modern relative to Apple's current CPU power management strategy, microprocessor to enter and return from a low-power mode since the CPU and associated logic are responsible for about are typically low. For instance, the AT&T Hobbit and certain 32% of power use; similar techniques should be applicable versions of the MC68030 and MC68040 use static logic so to operating systems with similar behavior. that most of their state can be retained when the clock is shut down [15]. Also, the PowerPC 603 can exit the low-power 1 Introduction Doze mode in about ten system clocks [4]. Because of the short delay and low energy cost for enter- Power consumption is becoming an increasingly impor- ing and leaving a low-power state, the optimal CPU power tant feature of personal computers. High power consumption management strategy is trivial: turn off the CPU whenever in desktop computers is undesirable as it leads to fan noise, there is no useful work to do. An opportunity for such a strat- heat, and expense. High power consumption in portable egy is described by Srivastava et al. [15], who point out that computers is even more undesirable as users of such ma- the process scheduling of modern window-based operating chines want them to last as long as possible on a single bat- systems is event-driven, i.e. that the responsibility of pro- tery charge. For these reasons, much work has been done in cesses in such systems is to process events such as mouse reducing the power consumption of computers. clicks when they occur and then to block until another such In [9] and [10], we analyzed the power consumption of event is ready. In this type of environment, the most appro- various Macintosh PowerBook computers, in typical use by priate strategy is to shut off the processor when all processes are blocked, and to turn the processor back on when an ex- This material is based upon work supported by a National Science Foundation Graduate Research Fellowship, by Apple Computer, and also ternal event occurs. An essentially equivalent version of this in part by the National Science Foundation under grants MIP-9116578 and strategy, namely to establish a virtual lowest-priority process CCR-9117028, the State of California under the MICRO program, Intel Cor- whose job is to turn off the processor when it runs, is recom- poration, Sun Microsystems, Fujitsu Microelectronics, Toshiba Corpora- mended by Suessmith and Paap [16] for the PowerPC 603, tion, and Sony Research Laboratories. Although the work presented here has bene®ted from support from and cooperation by Apple, the relation- and by Suzuki and Uno [17] in a 1993 patent. Such a virtual ship between this work and Apple's product development plans has yet to lowest-priority process has in the past been called the ªidle be determined. loop,º and in mainframes typically lighted a bulb on the con- ergy savings and performance is obtained. We call these lat- sole. ter two techniques the greediness technique and sleep exten- sion technique, respectively. We will show how using these 1.1 Why it isn't trivial techniques can improve the effectiveness of the basic strat- egy, allowing it to far surpass the effectiveness of the current We refer to the strategy of turning off the processor when MacOS inactivity-timer based strategy. Each of these is de- no process is available to run the basic strategy. Unfortu- scribed in more detail below. nately, in Apple's MacOS, processes can run or be scheduled In this paper, we evaluate these different strategies, over to run even when they have no useful work to do. This fea- a variety of parameter values, using trace-driven simulation. ture is partially by design, since in a single-user system there These simulations enable us to compare these algorithms to is less need for the operating system to act as an arbiter of re- the current MacOS strategy, and to optimize their parame- source use [12]. Partially, it is because the OS was not writ- ters. A comparison between two strategies is based on two ten with power management in mind. Partially, it is because consequences of each strategy: how much processor power the MacOS, like other personal computer operating systems it saves and how much it decreases observed performance. (e.g. those from Microsoft), is based on code originally de- The paper is structured as follows. In Section 2, we de- veloped for 8- and 16-bit non-portable machines, for which scribe in detail the strategies we will be comparing, including development time and code compactness were far more im- the current strategy used by MacOS, the basic strategy, and portant goals than clean design or faithfulness to OS design our suggested process management techniques for improving principles as described in textbooks. the basic strategy. In Section 3, we describe the methodology There are two main problems with the current manage- we used to evaluate these strategies: the evaluation criteria, ment of processor time, one having to do with the system the tools we used for the trace-driven simulation, and the na- and one having to do with applications. The ®rst problem is ture of the traces we collected. In Section 4, we present the that the operating system will sometimes schedule a process results of our simulations. In Section 5, we discuss the mean- even though it has no work to do. We were ®rst made aware ing and consequences of these results, and point the way to of this phenomenon when we studied traces of MacOS pro- future work. cess scheduling calls, and found that often a process would be scheduled to run before the conditions the process had 2 Strategies established as necessary for it to be ready were ful®lled. It seems that often, when there are no ready processes, the OS picks one to run anyway, usually the process associated with 2.1 Current strategy the active window. The second problem is that programmers The currently used processor power management strategy writing applications generally assume that when their appli- in MacOS is based on an inactivity timer. The operating sys- cation is running in the foreground, it is justi®ed in taking as tem will initiate processor power reduction whenever no ac- much processing time as it wants. First, a process will often tivity has occurred in the last two seconds and no I/O activity request processor time even when it has nothing to do. We has occurred in the last 15 seconds. Power reduction is halted discovered this problem in MacOS when we discovered pe- whenever activity is once again detected. Activity is de®ned riods of as long as ten minutes during which a process never here and in later contexts as any user input, any I/O device did anything, yet never blocked; we describe later what we read or write, any change in the appearance of the cursor, or mean by ªnever did anything.º Second, when a process de- any time spent with the cursor as a watch. The reason for the cides to block, it often requests a shorter sleep period than classi®cation of these latter two as activity is that MacOS hu- necessary.